Anthropic warns it could miss a self-improvement milestone

Anthropic has warned in updated IPO filings that its AI systems could imminently achieve recursive self-improvement. The company said it might recognise the milestone only after it has occurred because it relies on lagging indicators to track the pace of development.

The warning highlights a potential gap between advancing model capabilities and the safeguards intended to control them. Anthropic cautioned that if capabilities outpace safety, alignment and security measures, it could fail to mitigate catastrophic risks to humanity.

The company committed to maintaining human supervision over frontier model development, including delaying releases of powerful models such as Mythos. Investors were also warned that these precautions could carry commercial consequences: additional safety measures could increase operating costs and leave Anthropic at a competitive disadvantage.